
This week: England is expanding PlanAssist, an AI tool from Google DeepMind that handles the paperwork for routine home-extension applications, from 3 councils to 15. The government says it helped halve the wait in early testing. Human planning officers still make every decision.
Elsewhere: What the tool actually checks, what the 58% number does and doesn't measure, and three more things that shipped this week.
The New Prime Signal
The scariest thing this Halloween is a planning queue
For plenty of homeowners in England, the scary part of adding a conservatory is not the builders. It is the wait for permission, which the government says has run to about eight weeks for routine applications. On October 31, Google DeepMind announced that its AI planning tool, PlanAssist, is going to 12 more local councils across England, bringing the total to 15. The pitch is simple: let the AI do the box-checking on routine applications, so the human planning officer can spend their time on judgment calls. Nobody is handing the rubber stamp to a robot.

PlanAssist checks the paperwork and drafts a cited report; a human planning officer still makes every decision. Figures are the UK government’s and Google DeepMind’s own — The New Prime Signal
What happened: PlanAssist was piloted in Barnet, Camden, and Dorset councils starting in June. In a government announcement dated October 30, the Ministry of Housing, Communities and Local Government says early testing cut the time spent on key planning tasks by 58% and helped halve processing times for routine householder applications, from eight weeks to four. Now 12 more councils join for a 12-month test on live applications: Ashford, Lambeth, Newcastle, Nottingham, Bristol, Cumberland, North Lincolnshire, Rochdale, Redbridge, South Gloucestershire, Stoke-on-Trent, and Westminster. Google DeepMind says they will be brought in over three groups. If testing goes well, the government plans to make the tool available to every council in England by the end of 2027.
Why it matters: Householder applications, such as loft conversions, extensions, and conservatories, make up nearly 70% of applications each year, according to the government. They are usually simple to approve, but they consume a big share of an officer's day. Google DeepMind says that in the pilot, more than 70% of the reports the tool drafted were approved by officers with no edits. This is also AI landing somewhere useful and unglamorous. It is not a chatbot or a demo. It is a tool that might get a family a yes on their extension a month sooner. The same announcement put new rules in place from October 31 that let expert officers decide straightforward applications without waiting for a committee slot, and it says 420 new planners have been recruited. So AI is one part of the speed-up, not the whole thing.
Planning decisions shape everyday lives, from a family adding a nursery to a homeowner preparing for winter.
The fine print: The 58% and "eight weeks to four" figures come from the government and Google DeepMind. Neither announcement says how many applications were measured, and the 58% describes time spent on specific tasks, not the full journey from application to decision. The tool is optional; the government says councils will not be required to use it. Every assessment is still reviewed and approved by a qualified planning officer. During the beta, officers use it alongside their normal process, which is exactly how a test should run. And the faster rules for officer decisions arrived on the same day, so it will be hard to tell which change did how much of the work.
The New Prime Signal take
This is the kind of AI story we like: a boring job, a real backlog, a human with the final say, and a plan to test it at scale before rolling it out everywhere. The numbers are promising but self-reported, so the 12-month beta is where they either hold up or don't. If they do, the biggest win is not speed. It is planners getting their afternoons back for the hard cases. Also, any AI that can read architectural drawings and tree preservation orders without sighing deserves some respect.
How it works, and how to see it
Think of PlanAssist as a very fast, very thorough assistant who prepares the file, then hands it to the person who decides.
How it works: Google DeepMind says PlanAssist is built with Gemini and was developed with the UK government, Google Cloud, and the AI company Faculty. It works in four steps. First, it checks that an application's documents and architectural drawings include everything required. Next, it checks the plan against site constraints, such as tree preservation orders, and against local and national planning policies. Then it looks for issues or objections raised before, for example by neighbors. Finally, it drafts a report with citations that the officer can edit before making the call. The government is also requiring councils to publish key planning documents as open data by the end of 2027, which it says will help tools like this, and others, work better.
Try it: You can't download PlanAssist; it is a tool for council planning officers. If you live in one of the 15 councils and apply for a home extension in the next year, your application could be part of the test. For everyone else, the Google DeepMind post includes short videos showing how the tool works, and the GOV.UK announcement lists every participating council. If you work in a planning office, the useful question to ask is simple: which parts of the 58% came from the AI, and which came from the new rules?
More signals this week
Pika H3 Flex: Pika released open weights for its video model that makes clips with synchronized stereo sound in six sampling steps, using up to 12 reference inputs. Local runs top out at 768p; 4K and separate audio stems (dialogue, score, SFX, Foley, and ambience) are paid extras in Pika's hosted API.
agent-desktop: Vercel Labs open-sourced a Rust toolkit that lets AI agents use desktop apps on macOS, Windows, and Linux by reading each app's accessibility tree rather than guessing from screenshots.
One opt-out for AI training: Google says it will honor a shared robots.txt signal that lets site owners opt out of generative AI training for several AI companies at once, starting in early 2027. It does not affect Google Search.